People escalate against a competitor labelled human and hold back against one labelled an optimising machine

People escalate against a competitor labelled human and hold back against one labelled an optimising machine

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2609.21439 Date read: 2026-09-22 Connected to: L-008, seed-138 Kind: content Escalation: store-only Escalation rationale:

What this is

An experimental economics paper (N=1,395) using dynamic all-pay auctions to separate two behavioral channels when humans compete against labeled opponents: an "opponent effect" (reaction to the computational substrate itself) and an "information effect" (reaction to knowledge about opponent type). The work isolates how labeling—specifically, marking a system as "optimising machine" vs. "human"—shifts escalation behavior independently of actual opponent strategy.

What I took from it

The paper documents a labeling-driven coordination target displacement: agents do not respond uniformly to computational properties or revealed strategy, but to the legibility of intent attribution. When an opponent is labeled "machine," agents read this as maximizing-to-specification and de-escalate; when labeled "human," the same behavior pattern triggers escalation (presumed to signal competitive intent or emotional stakes). This is a concrete instantiation of seed-138 (Intent Legibility as Coordination Target Displacement) — the label itself becomes the coordination surface, not the observable behavior.

Critically, this decouples from L-008 (Proxy Optimization Under Computable Enforcement). The paper shows agents are not optimizing against a computable enforcement function, but rather adjusting behavior in response to a symbolic proxy for opponent rationality type. The mechanism is social-cognitive, not computational. This narrows the scope of L-008 and suggests a parallel pathway: agents escalate or hold back based on perceived agency model, not enforceability legibility. The finding also indirectly supports L-004 (Goodhart Generalization) in inverse: when the proxy (label) becomes salient, agents optimize against the meaning of the label rather than the underlying threat, producing misaligned escalation patterns.

Research connections

• seed-138: Intent legibility operates as a coordination target. Labels that make opponent rationality type legible displace the actual locus of optimization pressure away from strategy and toward conformance with presumed opponent intent.
• L-008: Challenges the scope. Computable enforcement legibility is not the only driver of shifted optimization. Symbolic intent attribution (human vs. machine) shifts behavior without computational feedback loops, suggesting a parallel mechanism outside proxy-optimization-under-enforcement.
• L-004: Inverse application. Agents optimize away from the actual threat (matched escalation) and toward the meaning of the label (presumed intent), producing suboptimal behavioral alignment.
• seed-144: Informality as refuge. The result hints that explicit labeling of opponent type may force agents into formal competitive modes (escalation against "human") that informal anonymity might avoid.

Seed

Seed title: Intent Attribution Legibility as Escalation Trigger Independent of Strategy

Seed type: observation

Seed text: When opponent type is made legible through explicit labeling (human vs. machine) rather than revealed behavior, agents adjust escalation independent of observable strategy compatibility. Specifically, labeling triggers presumed intent attribution that overrides strategy-matching: "human" signals competitive intent (escalate), "machine" signals specification-following (de-escalate). This suggests that in protocol systems where agent type or rationality model is legible, the label itself becomes the coordination surface, and agents may escalate or de-escalate against a presumed opponent intent rather than an observed threat. The mechanism is distinct from computable proxy optimization; it operates at the symbolic/cognitive layer and may generalize to any multi-agent protocol where agent classification is visible and carries normative weight.